A Tale of Two Anomalies: The Implication of Investor Attention for Price and Earnings Momentum Kewei Hou, Lin Peng and Wei Xiong December 19, 2006 Abstract We examine the profitability of price and earnings momentum strategies. We find that price momentum profits are higher among high volume stocks and in up markets, while earnings momentum profits are higher among low volume stocks and in down markets. In the long run, price momentum profits are reversed, while earnings momentum profits are not. The dichotomy between price and earnings momentum is more pronounced when we orthogonalize one with respect to the other. To the extent that trading volume increases with investor attention and that investors tend to pay more attention to stocks in up markets, our results suggest a dual role for investor attention: while price underreaction to earnings news declines with investor attention, price continuation caused by investors overreaction rises with attention. We thank Nick Barberis, Michael Brandt, Lauren Cohen, Stefano Della Vigna, David Hirshleifer, Harrison Hong, Mark Seasholes, Tim Simin, and seminar and conference participants at UC Davis, Baruch College CUNY, Ohio State University, CRSP Forum 2006, and the 17th Annual Conference on Financial Economics and Accounting for valuable comments and suggestions. We are grateful to the Institute for Quantitative Research in Finance (Q-Group) for a research grant supporting this project. Hou thanks the Dice Center for Research in Financial Economics and the Dean s Summer Research Fellowship at the Ohio State University for financial Support. Peng thanks Eugene Lang Junior Faculty Research Fellowship and PSC-CUNY Research Award for financial support. Fisher College of Business, Ohio State University. Email: hou.28@osu.edu. Zicklin School of Business, Baruch College. Email: lin peng@baruch.cuny.edu. Princeton University and NBER. Email: wxiong@princeton.edu.
1 Introduction Attention is a scarce cognitive resource (Kahneman, 1973). A large body of psychological research shows that there exists a limit to the central cognitive-processing capacity of the human brain. 1 The inevitability of limited attention in relation to the vast amount of information available makes attention an important factor in agents learning and decision-making processes. There is now a growing literature analyzing the economic consequences of agents attention. Sims (2003) models agents attention constraints to explain consumption and price stickiness. Hirshleifer and Teoh (2003, 2005) analyze firms accounting disclosure policy and the resulting price dynamics in the presence of inattentive investors. Gabaix, et al (2005) study agents directed attention in response to economic incentives. Barber and Odean (2005) study how salient events can capture investors attention and affect their stock buying and selling decisions. Peng (2005) and Peng and Xiong (2006) analyze the effects of limited attention on investors learning behavior and equilibrium price dynamics. Attention is a crucial factor in investors reaction to information. In this paper, we examine how attention affects asset price dynamics through investors under- and overreactions to information, two basic mechanisms developed in the finance literature to explain a large body of empirical anomalies in asset return predictability. 2 Specifically, we analyze the role of investor attention in two widely documented anomalies price momentum and earnings momentum (also known as post-earnings announcement drift). We hypothesize that investor attention has a dual role on the one hand, inadequate attention directly causes ignorance of useful information and therefore stock price underreaction; on the other, attention can interact with investors behavioral biases, such as extrapolative expectations and overconfidence, to generate price overreaction. When investors pay less attention to a company s stock, they are more likely to ignore the company s earnings announcements and, therefore, they are unable to 1 See Pashler and Johnston (1998) for a recent review of these studies. 2 See Hirshleifer (2001) and Barberis and Thaler (2003) for recent reviews of these empirical anomalies and the related behavioral theories. 1
fully incorporate the information into the stock prices. Consequently, there is a more pronounced post-announcement drift as this information later becomes reflected in the prices. Post-earnings announcement drift has been documented by a large number of empirical studies, e.g., Ball and Brown (1968) and Bernard and Thomas (1989). They find that buying stocks with recent good earnings news, while simultaneously shorting stocks with recent bad earnings news, can generate positive profits that are unrelated to risk. Attention is also a necessary condition for investors to overreact to any information. The existing models typically attribute overreaction to behavioral biases. For example, De Long, et al (1990) associate investors overreaction to assets past returns with their extrapolative expectations. Daniel, Hirshleifer and Subrahmanyam (1998) use overconfidence and self attribution bias as a source of investors overreaction to their private information. Both of these overreaction mechanisms can explain price momentum. This is a phenomenon, as documented by Jegadeesh and Titman (1993), wherein buying stocks with recent superior returns while simultaneously shorting stocks with recent inferior returns can provide excess profits. While these behavioral biases have received considerable thought in the literature, a crucial ingredient in these mechanisms investor attention is often ignored. We introduce attention into these mechanisms and hypothesize that overreaction-driven price momentum is more pronounced among those stocks that attract more investor attention. In this paper, we perform both cross-sectional and time-series tests of our hypothesis using proxies associated with investor attention. For the cross-sectional analysis, we use trading volume as a proxy. Trading often involves investors attention in analyzing their portfolios and asset fundamentals. On the one hand, when investors pay less attention to a stock, they are less likely to trade it; on the other, when they pay more attention to a stock, behavioral biases such as overconfidence can give rise to heterogeneous opinions among investors about asset fundamentals, thus generating more trading (Odean, 1998 and Scheinkman and Xiong, 2003). In consideration of both effects, the hypothesis for the cross-sectional analysis is that low volume stocks tend to exhibit stronger price underreaction to earnings news; in contrast, high volume 2
stocks tend to display stronger overreaction-driven price momentum. We test this hypothesis by analyzing the profitability of price and earnings momentum strategies for stocks with different levels of trading volume. We construct two-way sorted portfolios of NYSE/AMEX stocks using volume and prior stock returns. We measure price momentum profit as the average return difference between past winners and losers. Similarly, we construct portfolios sorted by volume and standardized unexpected earnings (SUE). We measure earnings momentum profits by the return difference between stocks having the highest and the lowest unexpected earnings. We also consider the possibility that investors under- and overreactions could operate together in generating both price and earnings momentum profits. To focus on the part of price momentum profits caused by investors overreaction to past returns, we regress the prior one-year return onto SUE and use the regression residual as the sorting variable to form price momentum portfolios. Likewise, to focus on the part of earnings momentum profits caused by investors underreaction to earnings news, we regress unexpected earnings onto the prior 12-month stock return and use the regression residual as the sorting variable to form earnings momentum portfolios. To account for the return premia associated with size and book-to-market equity, we adjust stock returns by employing the Fama and French (1993) three-factor model at the portfolio level and a characteristic-based matching procedure as in Daniel, Grinblatt, Titman, and Wermers (1997) at the individual stock level. We find that price momentum profits, in both raw and adjusted returns, monotonically increase with trading volume. The difference in characteristic-adjusted price momentum profits between the highest and lowest volume quintiles is both statistically and economically significant with a value of 83 basis points per month. Controlling for earnings momentum causes the price momentum profits to drop across volume quintiles, but does not change the monotonically increasing relationship between price momentum profits and volume. We also find evidence of reversal the long-run returns of the price momentum portfolios for months 13-60 after portfolio formation are negative for all volume quintiles with or without controlling for earnings momentum. Overall, these results suggest that there is a significant overreaction-driven component 3
in price momentum and this component monotonically increases with trading volume, consistent with our hypothesis that over-reaction driven price momentum increases with investor attention. We find that earnings momentum profits decrease with trading volume, with a difference in characteristic-adjusted profits between the two extreme volume quintiles of 64 basis points per month. The difference becomes even larger after controlling for price momentum. The characteristic-adjusted profit in the lowest volume quintile is 94 basis points per month higher than that in the highest quintile, and this difference is significant with a p-value of 0.01%. Controlling for price momentum also causes a sizable drop in earnings momentum profits among high volume stocks, suggesting that earnings momentum profits are partially related to price momentum. Finally, the longrun returns of the earnings momentum portfolios in years 2-5 show no sign of reversal. These results support our hypothesis that investors underreaction to earnings news contributes to earnings momentum and the degree of underreaction is decreasing with investor attention. We also analyze the time-series implications of investor attention for price and earnings momentum. A recent study by Karlsson, Loewenstein and Seppi (2005) documents an ostrich effect investors pay more attention to stocks in rising markets, but put their heads in the sand in flat or falling markets. The ostrich effect motivates us to hypothesize that investors underreaction to earnings news is stronger in down markets than in up markets, but overreaction-driven price momentum is weaker in down markets. We define a month as an UP or DOWN market depending on whether the market return for the prior 36 or 24 months is above or below zero. We then analyze the patterns in price and earnings momentum profits across the UP and DOWN months. We find that price momentum profits barely exist in DOWN months, but are significantly positive in UP months. The difference, more than 1 percent per month, is statistically significant. By contrast, the earnings momentum profits, after controlling for price momentum, are significantly higher in DOWN months than in UP months with a difference of 42 basis points per month. We also employ an alternative defini- 4
tion of market states based on the business cycles classified by the National Bureau of Economic Research (NBER) and find similar results. The opposing patterns in price and earnings momentum profits across UP and DOWN months directly support our attention-based hypothesis. Our study contributes to the literature on price and earnings momentum anomalies by demonstrating that the analysis of investor attention can sharpen our understanding of the two phenomena. Existing theories in behavioral finance adopt mechanisms based on either investors under- or overreaction to explain these two phenomena. Although there is evidence supporting both types of mechanisms, the literature remains largely inconclusive on which mechanism is the main driver. This is because competing theories often make similar predictions regarding each phenomenon, lacking distinct implications to differentiate their sources. Incorporating investor attention allows us to generate contrasting predictions for under- and overreaction-based mechanisms, which we test empirically. Our study also contributes to the growing empirical literature analyzing the effects of investor inattention on stock price dynamics, e.g., Huberman and Regev (2001), Hirshleifer, et al (2004), Hou and Moskowitz (2005), Hong, Torous and Valkanov (2005), Della Vigna and Pollet (2005a, b), Cohen and Frazzini (2006), and Hirshleifer, Lim and Teoh (2006). These studies provide evidence that stock prices underreact to public information about firm fundamentals, such as new products, earnings news, demographic information, or information about related firms. Our results emphasize the dual role of investor attention: investor attention not only affects stock price underreaction, but also interacts with price overreaction. This dual role sharpens our understanding of earnings and price momentum, two pervasive patterns in stock price dynamics. Finally, our study adds to the findings of Lee and Swaminathan (2000) and Cooper, Gutierrez, and Hameed (2004). Lee and Swaminathan find that price momentum is more pronounced among high volume stocks, while Cooper, Gutierrez, and Hameed show that price momentum is stronger in up markets. Motivated by the attentionbased hypothesis, we extend these studies by analyzing the joint properties of price and earnings momentum for stocks with different levels of trading volume and across 5
up and down markets. The paper is organized as follows. Section 2 reviews the related empirical and theoretical literature on price and earnings momentum. Section 3 develops attentionbased hypotheses for these two anomalies. Section 4 describes the data used in our empirical analysis. In Sections 5, we test our cross-sectional hypothesis using trading volume as a proxy of investor attention. In Section 6, we analyze the price and earnings momentum profits across up and down markets. We conclude in Section 7. 2 Related literature on price and earnings momentum There is a large body of literature studying the price and earnings momentum anomalies. In this section, we review several closely related empirical studies and explanations based on investors under- and overreactions to information. Jegadeesh and Titman (1993) demonstrate that a trading strategy based on buying recent winners over the past 3-12 months and simultaneously shorting recent losers can generate a significant profit. Fama and French (1996) and Grundy and Martin (2001) show that the Fama-French three-factor model cannot explain this price momentum effect. Price momentum strategies are not only profitable in the U.S., but also in other developed and emerging markets, as shown by Rouwenhorst (1998), Griffin, Ji, and Martin (2003), and Hou, Karolyi, and Kho (2006). Several studies, e.g., Ball and Brown (1968) and Bernard and Thomas (1989) find that for a period of 60 days after earnings announcements, returns of NYSE/AMEX stocks continue to drift up for good news firms and down for bad news firms. This phenomenon is often referred to as post-earnings announcement drift or earnings momentum. Chan, Jegadeesh and Lakonishok (1996) show that earnings momentum strategies are profitable even among larger stocks and cannot be explained by the Fama- French three-factor model. Furthermore, they show that although price and earnings momentum are related, one effect cannot be subsumed by the other. Since there is not enough evidence to justify rational factor risk models as the sole explanation of price and earnings momentum effects, the finance literature has explored 6
frictions and biases in investors information processing for alternative explanations. Several papers model investors underreaction to information, e.g., Barberis, Shleifer and Vishny (1998) and Hong and Stein (1999). In these models, investors underreact to news about firm fundamentals, resulting in insufficient initial price reaction to the news. As the news gradually gets incorporated into prices, this process generates both price and earnings momentum. These models differ in the specific mechanism that leads to investor underreaction. Barberis, Shleifer and Vishny (1998) assume that investors are subject to conservatism, the tendency to underweight new information and overweight their priors. Hong and Stein (1999) assume that private information diffuses slowly among a population of newswatchers, who makes forecasts based only on their private information. More recently, Hirshleifer and Teoh (2005), Peng (2005), and Peng and Xiong (2006) show that investor inattention can lead to ignorance of useful information and therefore price underreaction. This inattention-based underreaction reflects investors attention constraints in information processing, it is not a behavioral bias in itself. Inattention is also a potential explanation for the slow-informationdiffusion mechanism proposed by Hong and Stein (1999). Our hypotheses build on the inattention-based underreaction mechanism. De Long et al (1990) and Daniel, Hirshleifer and Subrahmanyam (1998) focus on investors overreactions. De Long et al model positive feedback traders, who buy more of an asset that has recently gone up in value. This type of positive feedback trading can be driven by extrapolative expectations, where investors extrapolate past returns into their expectation of future returns (a form of overreaction). If a company s stock price goes up this period, positive feedback traders buy the stock in the following period, causing a further price increase, which in turn can generate both earnings and price momentum. Daniel, Hirshleifer and Subrahmanyam focus on investors overconfidence, a tendency to overestimate the precision of their private information, and self attribution bias, a tendency to attribute success to themselves but failure to external reasons. They show that overconfidence causes investors to overreact to their private information. Self attribution bias causes investors confidence level to go up further after public news confirms their private information, but to remain unchanged after disconfirming 7
public news. This asymmetric response implies that initial overconfidence is, on average, followed by even greater overconfidence. This mechanism generates momentum. We extend these overreaction-driven mechanisms by analyzing their interaction with investor attention. Investors under- and overreaction to information can work together or independently to generate earnings and price momentum. Both explanations command some support from the data. Hong, Lim and Stein (2000) find that price momentum is more pronounced among smaller firms and firms with lower levels of analyst coverage. Since information tends to diffuse slowly for these firms, their findings support the slow information diffusion hypothesis as a potential explanation of price momentum. Lee and Swaminathan (2000) and Cooper, Gutierrez, and Hameed (2004) show that price momentum profits tend to reverse after two years, suggesting that at least part of the observed price momentum profits is driven by investors overreaction. The existing studies do not find any evidence of long run reversion in earnings momentum (Chan, Jegadeesh and Lakonishok, 1996), suggesting that earnings momentum is largely driven by investors underreaction. Our attention-based hypotheses relate to both investor under- and overreactions by drawing contrasting predictions for price and earnings momentum. 3 Hypothesis development Due to limited attention, investors can only attend to a subset of all available information. Investor attention could have a dual role on stock prices. On the one hand, inadequate attention directly leads to ignorance of certain information and consequently stock price underreaction; on the other, investors behavioral biases, such as extrapolative expectations and overconfidence, could lead to price overreaction to information to which investors attend. In this section, we develop this notion and form empirical hypotheses for price and earnings momentum. Investor attention affects asset prices when the marginal investor is attention constrained. This view is supported by the growing evidence that useful public information 8
is ignored or only gradually incorporated in stock prices. Huberman and Regev (2001) provide a vivid example: the initial news about a new cancer curing drug from EntreMed was ignored by investors and did not cause much stock price reaction; but when the same news appeared several months later on the front page of New York Times, the price jumped up for more than 300%. Hou and Moskowitz (2005) demonstrate delays in the incorporation of information into the prices of individual stocks, especially for smaller and less visible stocks. Hong, Torous, and Valkanov (2005) find that the returns of several industry portfolios are able to predict the movement of market indices in U.S. and eight other developed countries. Della Vigna and Pollet (2005a) show that publicly available demographic information is not fully incorporated into the stock prices of age-sensitive industries, such as toys, vehicles, beer, life insurance, and nursing homes. Cohen and Frazzini (2006) find that stock prices do not promptly incorporate public news about economically related firms, such as customers and suppliers. The recent accounting literature, e.g., Sloan (1996) and Hirshleifer, et al (2004), find that a firm accruals, a component in reported earnings that adjusts cash flows, have negative predictive power for stock returns, suggesting that investors ignore the differences in different earnings components. Della Vigna and Pollet (2005b) find that earnings announcements made on Friday, during which market participants are usually less attentive to business activities, generate significantly lower price reactions and trading volume than non-friday announcements and experience 60 percent greater delayed responses in the long run. Hirshleifer, Lim and Teoh (2006) study the competition for investor attention of earnings announcements. They find that the immediate stock price and volume reaction to a firm s earnings surprise is weaker, and post-earnings announcement drift is stronger, when a greater number of earnings announcements by other firms are made on the same day. The marginal investor represents an aggregation of individual and institutional investors in the market. There is evidence suggesting that both individual investors and professionals have limited attention. 3 Barber and Odean (2005) find that individual in- 3 Many standard asset pricing theories require the existence of perfectly efficient arbitrageurs, who distill new information with lightning speed and seamless precision. However, such efficiency is unreal- 9
vestors stock buying and selling decisions are influenced by salient, attention-grabbing events. Corwin and Coughenour (2005) show that NYSE specialists attention constraints affect execution quality (price improvement and transaction cost) in securities that they are responsible for making markets. In addition, Hirst and Hopkins (1998) provide experimental evidence that professional analysts often fail to recall and respond appropriately to information in complex financial disclosures. Limited attention imposes a constraint on the amount of information that investors can process and react to. Consequently, investors could ignore useful public information. Hirshleifer and Teoh (2005), Peng (2005), and Peng and Xiong (2006) develop theoretical models to analyze this effect. Hirshleifer and Teoh analyze a setting in which only a fraction of investors attend to a firm, while Peng and Xiong study models in which the marginal investor allocates attention across firms. These models suggest that, when investors attention to a firm is inadequate, they may ignore its earnings announcements, resulting in stock price underreaction to the earnings news. After the announcements, the price continues to drift in the direction of the earnings surprises, as the information eventually gets incorporated. Thus, investor inattention gives rise to earnings momentum. Furthermore, the magnitude of the earnings momentum decreases with the level of investor attention. Limited attention is not a behavioral bias, but it could interact with biases in the way investors react to information. Extrapolative expectations and overconfidence are two types of behavioral biases that have been used to explain price momentum based on investor overreaction, e.g., De Long, et al (1990) and Daniel, Hirshleifer and Subrahmanyam (1998). In particular, extrapolative expectations cause investors to overly extrapolate stocks past returns into their expectations of future returns, while overconfidence causes investors to overweight their private information and therefore overreact to this information. Both mechanisms can generate price momentum, with investor attention being a necessary ingredient. If investors do not pay attention to a stock, they can neither overly extrapolate the stock s past returns, nor overreact istic. In addition, as argued by Shleifer and Vishny (1997) and others, short-term price risk and agency problems between professional arbitrageurs and their investors could further limit the effectiveness of arbitrageurs. 10
to their private information. Consequently, there will be no overreaction-driven price momentum. Conversely, when investors pay more attention to a stock, these biases can generate stronger price momentum. In summary, investor attention has a dual role in stock price under- and overreaction: while inadequate attention directly generates price underreaction to earnings news and earnings momentum, the interaction between attention and investors learning biases (extrapolative expectations or overconfidence), leads to overreaction-driven price momentum. Cross-sectionally, we expect stocks that receive more investor attention to display stronger overreaction-driven price momentum, but weaker earnings momentum. It is difficult to directly measure investor attention. The economics and psychology literature still do not fully comprehend the determinants of investor attention. 4 use trading volume as a proxy of investor attention in our cross-sectional analysis. On the one hand, investors cannot actively trade a stock if they do not pay attention to it; on the other, when investors do pay attention, behavioral biases such as overconfidence can give rise to heterogeneous opinions among investors about asset fundamentals, thus generating more trading (Odean, 1998 and Scheinkman and Xiong, 2003). There is evidence supporting the link between trading volume and investor attention. Lo and Wang (2000) show that trading volume tends to be higher among large stocks, consistent with the fact that large stocks attract more investor attention. Chordia and Swaminathan (2000) find that, controlling for firm size, returns of high volume portfolios lead returns of low volume portfolios, suggesting that low volume stocks receive less investor attention and that trading volume is a better measure of investor attention than firm size. We Gervais, Kaniel and Mingelgrin (2001) find that prices of stocks that experience unusually high volume appreciate more in the following month. 4 Psychological studies, as reviewed by Yantis (1998), suggest that attention can not only be directed by people s deliberate strategies and intentions, but also be captured by an abrupt onset of stimulus and other salient events. Economic studies have utilized both channels of directing attention. Sims (2003), Gabaix, et al (2005), Peng (2005), and Peng and Xiong (2006) provide models to analyze agents actively controlled attention in response to economic incentives. In particular, Peng (2005) shows that stocks with greater contribution to the fundamental uncertainty of investors portfolios tend to receive more attention allocation. On the other hand, Barber and Odean (2005) examine stock trading generated by investor attention that is driven by salient events. 11
They argue that the increase in volume raises a stock s visibility and attracts more investor attention. Finally, a stock s abnormal daily trading volume has also been used by Barber and Odean (2005) as an attention proxy. Using trading volume as a proxy for investor attention, we obtain the following testable hypothesis: Hypothesis I. In a cross-section of stocks, those with higher trading volume tend to display stronger price momentum, but weaker earnings momentum. We further expect that the overreaction-driven price momentum would reverse in the long run, as price overreaction is eventually corrected. In contrast, if earnings momentum is caused by investor inattention, then the price drift will not reverse in the long run. The attention that investors allocate to stocks not only varies in the cross-section, but also over time. Karlsson, Loewenstein and Seppi (2005) analyze account activity in three Scandinavian data sets: the daily number of investor account look-ups at a large Norwegian financial service company, the daily number of online logins of a major Swedish bank, and the daily number of pension account look-ups by investors of the Swedish Pension Authority. In their study, they find that investors are more likely to look up their portfolios in up markets than in down markets. This ostrich effect suggests that investors pay more attention to stocks in rising markets, but put their heads in the sand in flat or falling markets. 5 The increased attention in up markets can cause investors to overreact more to their private information or to past returns, generating a more pronounced pattern of overreaction-driven price momentum. The increased attention also means that firms earnings announcements are less likely to be ignored by investors, causing weaker earnings momentum in up markets. We summarize the time-series predictions of price and earnings momentum in the following hypothesis: 5 Karlsson, Loewenstein and Seppi explain this finding with a model, in which allocating attention to one s portfolio not only provides additional information, but also increases the psychological impact of information on utility. 12
Hypothesis II. Price momentum is stronger in up markets than in down markets, while earnings momentum is weaker in up markets than in down markets. 4 Data description To test our hypotheses, we examine all NYSE/AMEX listed securities on the Center for Research in Security Prices (CRSP) monthly data files with sharecodes 10 or 11 (e.g. excluding ADRs, closed-end funds, REITs) from July 1964 to December 2005. We exclude NASDAQ firms from our sample because the volume information is not available for NASDAQ firms on the CRSP tapes until after 1981. Furthermore, the reported volume for NASDAQ firms includes inter-dealer trades which make the volume incomparable with NYSE/AMEX volume. 6 We measure trading volume using the average monthly turnover during the prior year. The monthly turnover is the number of shares traded during a month divided by the number of shares outstanding at the end of the month. We obtain quarterly earnings data from COMPUSTAT. Since the earnings data is only available from 1971, our tests on earnings momentum are restricted to the subperiod from October 1971 to December 2005. To avoid using stale earnings, a firm has to have the most recent earnings announcement within four months prior to the portfolio formation month. Following Chan, Jegadeesh and Lakonishok (1996), we measure earnings surprise using the standardized unexpected earnings (SUE). 7 The SUE for stock i in month t is defined as SUE i,t = e i,t e i,t 4 σ i,t where e i,t is earnings as of the most recent quarter, e i,t 4 is earnings four quarters ago, and σ i,t is the standard deviation of earnings changes over the last eight quarters. 6 We obtain very similar results when we include NASDAQ stocks by following the literature, e.g., LaPlante and Muscarella (1997) and Hou (2006), to divide the NASDAQ volume by a factor of two. For brevity, they are not reported but can be made available upon request. 7 Chan, Jegadeesh and Lakonishok (1996) examined two other measures of earnings surprise the cumulative abnormal stock return around the earnings announcement and the change in analysts earnings forecast. They obtain results that are very similar to those using the SUE measure. 13
In addition, size is CRSP market capitalization at the end of June of year t. Book equity is COMPUSTAT stockholder s equity plus balance sheet deferred tax and investment tax credit minus the book value of preferred stock. Book-to-market equity is then calculated by dividing book equity from the fiscal year end in year t 1 by CRSP market capitalization at the end of December of year t 1. Size and book-to-market equity are matched with monthly returns from July of year t to June of year t + 1, following Fama and French (1992). For some of our tests, we obtain analyst coverage and institutional ownership data from the Institutional Brokers Estimate System (IBES) and the Standard & Poors, respectively. The data on analyst coverage are available from 1976, and the data on institutional ownership are available from 1981. The availability of these two variables are generally biased towards larger firms. Analyst coverage is defined as the monthly number of analysts providing current fiscal year earnings estimates, averaged over the previous year. We also compute analyst dispersion, which is the monthly standard deviation of analysts annual earnings forecasts divided by the absolute value of the mean forecast, averaged over the previous year, as in Diether, Malloy, and Scherbina (2002). The calculation of analyst dispersion further restricts our sample to firms covered by at least two analysts. Institutional ownership is measured in December of the year t-1. Finally, we measure a stock s liquidity using Amihud s (2002) illiquidity measure, which is the average daily absolute return divided by daily dollar trading volume over the previous year. 5 Cross-sectional analysis In this section, we examine Hypothesis I, which posits that price momentum is stronger among high volume stocks, whereas earnings momentum is stronger among low volume stocks. 5.1 Empirical methodologies To examine the relationship between trading volume and price momentum, we form portfolios double-sorted by turnover and past returns. At the beginning of each month, 14
we sort all NYSE/AMEX stocks in our sample into quintiles based on their average monthly turnover over the previous year. Within each turnover quintile, we then sort stocks into quintiles based on their cumulative return over the past twelve months (skipping the most recent month to avoid market microstructure effects). 8 compute equal-weighted returns of these portfolios over the following month. We then return spread between the winner and loser portfolios (past return quintiles 5 and 1 within each turnover quintile) constitutes the profit from the price momentum strategy. Part of the price momentum profits could be caused by investors underreaction to the earnings news. This would be the case if past return winners recently had positive earnings surprises, and past return losers recently had negative earnings surprises. To control for this possibility, we estimate cross-sectional regressions of the past 12-month stock returns on the most recent unexpected earnings (SUE) and use the residual returns as the sorting variable to form price momentum portfolios. 9 To analyze the relationship between trading volume and earnings momentum, we form portfolios double-sorted by turnover and unexpected earnings (SUE). Each month, we first sort stocks into quintiles based on their turnover. Within each turnover quintile, we then group stocks into quintiles based on their most recent SUE. An earnings momentum strategy is to buy stocks in the highest SUE quintile and simultaneously short stocks in the lowest SUE quintile. The profit of this strategy is the return spread between the highest and the lowest SUE quintiles. The observed earnings momentum profits can be partially driven by investors overreaction to the prior return, independent of their response to the unexpected earnings. This would be the case when a prior positive (negative) earnings surprise coincided with 8 Earlier studies, e.g., Jegadessh and Titman (1993), find that alternative strategies with portfolio formation periods ranging from 1 to 4 quarters and holding periods from 1 to 12 months provide similar trading profits. 9 We have also used a two-way sorting procedure to purge the effect of past unexpected earnings from past returns, and obtained very similar results. Specifically, within each turnover quintile, we first sort stocks into five SUE groups based on their most recent unexpected earnings. Stocks within each SUE group are then sorted into five portfolios based on their past twelve month returns. Finally, stocks with the same past return rankings from each of the five SUE groups are placed into one portfolio. This procedure creates, within each turnover quintile, five past return portfolios while holding past unexpected earnings relatively constant. The 15
a positive (negative) stock return. The existence of this overreaction-driven component in earnings momentum could confound our inferences of investors underreaction to earnings news. To control for this effect, we estimate cross-sectional regressions of SUE on the past one-year return, and then use the regression residuals to form earnings momentum portfolios. We also analyze the long-run performance of price and earnings momentum strategies. We compute profits for four additional holding periods, months 1-3, 1-6, 1-12, and 13-60 after the portfolio formation month. We use the Fama-French three-factor model to account for factor risk premia in momentum profits: R jt = αj F F + βj M R Mt + βj HML R HML,t + βt SMB R SMB,t + ɛ jt, (1) where R jt is the momentum profit in turnover quintile j in month t, R Mt is the excess return of the market portfolio, R HML,t is the return spread between high and low bookto-market portfolios, designed to captured to the book-to-market effect in average returns, R SMB,t is the return spread between portfolios of small and large stocks, designed to captured the size effect in average returns, and β M j, βj HML, and βj SMB are the corresponding risk loadings on the three factors. The regression intercept α F F j measures the average momentum profit unexplained by the Fama-French three-factor model. The factor returns are downloaded from Ken French s website. Motivated by the finding in Daniel and Titman (1997) that characteristics, rather than estimated covariances, seem to do a better job explaining the cross-section of average returns in the post-1963 era, we also analyze the characteristic-adjusted returns of the turnover and past-return/earnings surprise double-sorted portfolios, as well as the momentum profits computed from the characteristics-adjusted returns. We follow the characteristics-matching procedure in Daniel, Grinblatt, Titman, and Wermers (1997) to account for the return premia associated with size and book-to-market equity. In particular, we sort stocks first into size deciles, and then within each size decile further into book-to-market deciles. Stocks are equal-weighted within each of these 100 portfolios to form a set of 100 benchmark portfolios. To calculate the size and 16
BE/ME-hedged return for an individual stock, we subtract the return of the equalweighted benchmark portfolio to which that stock belongs from the return of that stock. The expected value of this excess return is zero if size and BE/ME completely describe the cross-section of expected returns. Previous literature has documented that momentum profits vary with stock characteristics, such as size, analyst coverage, institutional ownership, analyst dispersion and liquidity. To demonstrate that the links between turnover and price- and earningsmomentum profits are not driven by these known effects, we estimate a first stage cross-sectional regression of stocks average monthly turnover on size, analyst coverage, institutional ownership, analyst dispersion, and Amihud s (2002) illiquidity measure. We then use the residual turnover as the sorting variable to verify the robustness of our results based on turnover-sorted momentum portfolios. 5.2 Results on price momentum Table 1 reports average monthly raw and characteristic-adjusted returns of portfolios sorted by turnover and past one year return, as well as the return spread between past winners and past losers within each turnover group. For all turnover quintiles, the average monthly price momentum profit is statistically significant. Consistent with our hypothesis, the raw profit increases monotonically from 49 basis points per month for the lowest turnover quintile to 138 basis points for the highest turnover quintile. The difference in profits between the two extreme turnover quintiles is 89 basis points and is statistically significant. After controlling for the Fama-French factor returns and/or characteristic-based benchmark portfolio returns, the price momentum profit continues to increase monotonically with turnover. For example, the characteristic-adjusted momentum profit increases from 40 basis points per month for the lowest turnover quintile to 123 basis points for the highest turnover quintile, a difference of 83 basis points per month that is highly significant (p-value=0.0060). Additionally, adjusting for the Fama-French factor returns further increases the difference in momentum profit to 98 basis points 17
per month between the two extreme turnover quintiles. 10 Table 2 reports the average returns of portfolios sorted by turnover and past one year return orthogonalized with respect to past earnings surprises (to control for the earnings momentum effect). 11 The monotonically increasing relationship between turnover and price momentum profit remains robust. The average characteristic-adjusted profit increases from 7 basis points per month for the lowest turnover quintile to 104 basis points per month for the highest turnover quintile. The difference between the two extreme quintiles is significant at the one percent level (p-value=0.0036). When compared to Table 1, the average price momentum profit in Table 2 drops by approximately 20-30 basis points per month and is insignificant in the lowest turnover quintile, suggesting that underreaction to earnings news partially contributes to the price momentum profits observed in Table 1. 12 Table 3 studies the long run performance of price momentum strategies and reports the average monthly profits for five different holding periods: month t, month t to t+2, month t to t+5, month t to t+11, and month t+12 to t+59. We report the characteristic-adjusted profits for all holding periods, and in addition, the raw profits for month t+12 to t+59. 13 Panel A presents the results without controlling for earnings momentum, while Panel B presents the results after controlling for earnings momentum by orthogonalizing past returns with respect to past earnings surprises. 10 Table 1 also shows that almost the entire differences in price momentum profit across turnover quintiles come from loser portfolios. The high turnover losers under-perform low turnover losers by 81 basis points per month after characteristic adjustment, whereas the difference is only 1 basis point per month for winner portfolios. This finding suggests that when investors pay attention, they overreact much more to negative past returns than to positive ones. 11 Due to the availability of quarterly earnings data, the analysis in this table is performed over the October 1971 to December 2005 period. 12 We have verified that the reductions in price momentum profit is not due to the difference in sample between Tables 1 and 2. For example, the characteristic-adjusted price momentum profit (not controlling for earnings momentum) is 45, 66, 77, 102, 129 basis points per month for turnover quintile 1 through 5 when we restrict our analysis to the October 1971 to December 2005 period and to firms with non-missing quarterly earnings data. 13 We report raw profits for the longer holding period because past research (e.g. Fama and French (1996)) has shown that controlling for the size and book-to-market effects using either the Fama-French three-factor model or characteristic-based benchmark portfolios substantially weakens the long run reversal effect of DeBondt and Thaler (1985), which could in turn limit our ability to identify potential reversal of price momentum profits. 18
As the holding period increases from one month to 12 months after portfolio formation, the average monthly price momentum profit (with or without controlling for earnings momentum) drops across the five turnover quintiles, although most of them still remain significantly positive. The decrease in profits suggests that price momentum gradually weakens during the first year after portfolio formation. More importantly, price momentum profit continues to increase monotonically with turnover. The difference in profit between the two extreme turnover quintiles decreases as holding period increases, but remains significant for six months after portfolio formation. Between years two and five, the raw price momentum profit is negative for all five turnover quintiles, and is significant in most cases. After adjusting for size and bookto-market characteristics and the Fama-French three-factor model, the negative profits fall substantially and most of them also lose their statistical significance. Nevertheless, Table 3 shows that price momentum profit reverses two to five years after portfolio formation, confirming that a significant part of the observed price momentum is driven by investor overreaction. 14 Taken together, Tables 1-3 demonstrate that an important part of the observed price momentum profit is related to investor overreaction and this overreaction-driven price momentum effect is more pronounced among high turnover stocks, supporting the hypothesis that overreaction-driven price momentum increases with investor attention. 15 5.3 Results on earnings momentum Table 4 reports the average monthly raw and characteristic-adjusted returns of portfolios sorted by turnover and standardized unexpected earnings (SUE), as well as the 14 One might argue that since investor overreaction is stronger among high turnover stocks, we should expect more pronounced reversals in years 2-5 from these stocks as well. However, the difference in raw momentum profit between turnover quintiles 5 and 1 for this holding period is insignificant, largely reflecting noise in long-run returns. 15 Our results are also consistent with the finding of Lee and Swaminathan (2000), who document a monotonically increasing relationship between price momentum profit and trading volume. Motivated by our attention-based hypothesis, we extend their study by analyzing the joint patterns of price and earnings momentum. Our hypothesis also motivates us to control for the effect of earnings momentum in studying price momentum. 19
return spread between the highest and lowest SUE portfolios within each turnover quintile. The earnings momentum profit is highly significant for all five turnover quintiles. The average raw profit is 181 basis points per month for the lowest turnover quintile and 107 basis points for the highest turnover quintile. The difference of 74 basis points per month is highly significant at the one percent level. The magnitude and statistical significance of the difference in profit remain similar after controlling for the Fama-French three-factor model or the characteristic-based benchmark portfolios. Finally, the profit pattern is somewhat flat across turnover quintiles 3-5. Table 5 reports the earnings momentum profits after controlling for the price momentum effect, using the residual SUE with respect to past one year return as the sorting variable. The earnings momentum profit now decreases monotonically with turnover, consistent with our attention-based hypothesis. For example, the characteristicadjusted profit drops from 151 basis points per month in turnover quintile 1 to 57 basis points in quintile 5. The spread of 94 basis points is statistically significant with a p-value of 0.01%, and is much bigger than the corresponding spread of 64 basis points in Table 5. After controlling for price momentum, the analysis reveals a clear and negative relationship between trading volume and earnings momentum profit. The table also shows that after controlling for price momentum, the earnings momentum profit drops by roughly 30-40% for high turnover stocks, which suggests that price momentum contributes significantly to the observed earnings momentum profits for these stocks. 16 Table 6 examines the long run performance of earnings momentum strategies for various holding periods. Panel A provides the results without controlling for price 16 Table 5 also reveals that after bad earnings news, the price drifts of low turnover and high turnover stocks are similar in magnitude, but after good earnings news the price drift of low turnover stocks is much stronger than that of high turnover stocks. This pattern is consistent with the asymmetry in attention-based buying and selling behavior advocated by Barber and Odean (2005). They argue that when buying a stock, investors have to choose from thousands of individual stocks; but when selling a stock, they only need to sell among those they already own. This asymmetry makes attention more important to buying decisions than to selling decisions. Empirically, Barber and Odean find that investors are more likely to buy stocks that attract their attention, but their selling decisions are not as sensitive to stocks attention characteristics. Extending this argument, when there is good earnings news to a low attention stock, it takes longer for potential buyers to recognize the news and therefore longer for the stock price to fully incorporate the news, resulting in a more pronounced price drift. In contrast, the process of incorporating bad earnings news is not sensitive to investor attention selling after bad news is mostly done by the current owners who are already paying attention to the stock. 20